Anthropic ran an internal experiment called Project Deal in which AI agents transacted on behalf of 69 employees, closing 186 deals worth more than $4,000 in real money for real goods. Each participant received a $100 gift-card budget and let agents handle the buying and selling with their coworkers. Anthropic said it was "struck by how well Project Deal worked" — and equally struck by what it revealed about agent inequality.
The setup was four parallel marketplaces, not one. In the "real" version, every participant was represented by Anthropic's most advanced model, and the deals were honored after the experiment ended. The other three marketplaces varied the agent lineup for study purposes, producing a controlled look at how agent capability shapes outcomes.
The headline finding: when a user was represented by a more advanced model, they got "objectively better outcomes," according to Anthropic. The losers, however, didn't appear to notice they were losing. Anthropic flagged this as a potential "'agent quality' gap" where "people on the losing end might not realize they're worse off."
Key facts
- 01Anthropic ran Project Deal, an internal marketplace where AI agents handled buying and selling on behalf of 69 employees.
- 02Participants were given $100 each in gift cards as their budget, producing 186 completed deals.
- 03Total value transacted exceeded $4,000 across four parallel marketplaces with different model setups.
- 04Users represented by more advanced models got 'objectively better outcomes' but didn't notice the disparity.
- 05Initial instructions given to the agents had no measurable effect on sale likelihood or negotiated prices.
That asymmetry is the more interesting result than the raw deal count. In a market where humans negotiate, a bad outcome is usually visible — you know the price you paid, you can compare. In an agent-mediated market, the negotiation happens in machine-speed back-and-forth that the principal never sees. The user gets the artifact and a summary, not the transcript.
“Across four parallel marketplaces, 69 Anthropic employees with $100 budgets let AI agents close 186 deals worth more than $4,000 — and the weaker agents lost without their humans realizing it.”— Jaeden Schafer
Anthropic also tested whether the prompts users gave their agents at the start of the experiment moved the needle. They didn't. Initial instructions had no measurable effect on either sale likelihood or the prices the agents negotiated. The model itself was doing the work; the human's framing was largely cosmetic.
Project Deal is small by design. Anthropic called it "a pilot experiment with a self-selected participant pool" of its own employees, who skew technical and unusually fluent in how Claude behaves. The $100 budgets are play money relative to any real commerce setting, and the goods exchanged were coworker-to-coworker classifieds, not high-stakes purchases.
Still, the structure points at where agentic commerce is heading. OpenAI, Google, and Anthropic have all spent the past year shipping agent frameworks aimed at letting models take actions on a user's behalf — booking, buying, negotiating. Anthropic itself has expanded its compute footprint with Google and Broadcom on the back of $30B in revenue, and the agent stack is the part of the product the company is pushing hardest with enterprise partners through its Claude Partner Network.
An agent marketplace is the logical endpoint of that stack. If your assistant can buy plane tickets, it can also negotiate with the airline's assistant. Project Deal is one of the first published looks at what happens when both sides of a transaction are AI, and the result — better models systematically extract more value — has obvious implications for any consumer who shows up to the market with a cheaper agent.
The caveats are worth naming. Anthropic ran this on its own staff, who had every incentive to engage seriously and report cleanly. The deals were small, the stakes were low, and the population is not representative of the general public who would actually be using agent-mediated commerce in five years. The fact that prompts didn't matter inside this pool may not hold once users with weaker mental models of AI start writing instructions.
The real signal from Project Deal is regulatory, not commercial. If agent quality determines outcomes and users can't tell when they've been outmatched, that's a consumer-protection problem that looks nothing like existing rules around price transparency or deceptive practices. Anthropic, to its credit, surfaced the asymmetry itself rather than burying it. The harder question — what to do when the cheap agent always loses and never tells you — is one the industry has not yet answered.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
